docs: algorithm, policy and contributing pages match the code

[tactino]

An audit of the algorithm, policy and contributing pages against plugrl-server found 21 discrepancies. This fixes them in both languages.

Wrong

• plugrl-run-server --help lists policies only. Algorithms appear under <policy> <variant> --help.
• Six algorithms are registered, not five. ppo was missing entirely, and now has a subsection: CleanRL's PPO, the default and dppo-square variants, and which policies it takes.
• No algorithm needs the dppo extra. dppo-policy and dppo-gaussian-policy do. Without it they are silently absent from the CLI; the quoted log line did not exist.
• pi0-policy does run under dppo, with the libero variant, as E25 did.
• --policy.train-expert-only true is a parse error. The pages now use the flag form.
• The E11 command now matches E11's script.

Stale. The training loop page now describes the server after #104 and #106:

• no inference while the algorithm wants to learn;
• frames from an older policy, or arriving while it waits to learn, go to discard_feedback and are counted in server/discarded_frames;
• global_step counts stored frames;
• the rule applies to on-policy algorithms, with on_policy = False for replay buffers.

The page also now covers resuming: the directory layout, FileNotFoundError and FileExistsError, and the Ctrl-C checkpoint. It covers warm starts from another run's weights (--algo.policy-checkpoint-path, --algo.restore) as well.

Missing

• Custom algorithm contract: discard_feedback (an override must call super()), on_policy, derive_train_state, pre_learn/post_learn, init_optimizers, and load_checkpoint on --resume. Registration needs both the config module and the class module imported.
  • The template now uses global_steps, and calls record_episode_metrics, without which rollout/* stays 0.
• Policies: gaussian-policy and dppo-gaussian-policy; batch-first (B, H, ...) actions; action_dim/action_horizon for the metadata; a runtime state that can be sliced per env; and the critic DPPO and FPO require.
• Flow policies: FPO's base, _predict_v and dt.
• InternalState no longer exists; the pages name the real types.

Checked. Every plugrl-run-server command on these pages parses against the CLI. pi0-policy was checked with its OpenPI imports stubbed out, as OpenPI is not installed here. The custom algorithm template was loaded and each method called on dummy-policy. mkdocs build --strict passes.

Merge after plugrl-server #106 (on_policy) and #107 (the dummy policy's defaults), which these pages describe.